CN108509834A - Graph structure stipulations method based on video features under polynary logarithm Gaussian Profile - Google Patents
Graph structure stipulations method based on video features under polynary logarithm Gaussian Profile Download PDFInfo
- Publication number
- CN108509834A CN108509834A CN201810048588.6A CN201810048588A CN108509834A CN 108509834 A CN108509834 A CN 108509834A CN 201810048588 A CN201810048588 A CN 201810048588A CN 108509834 A CN108509834 A CN 108509834A
- Authority
- CN
- China
- Prior art keywords
- vertex
- graph structure
- optical
- flow feature
- video
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
- G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/46—Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical fields of video features optimization, and in particular to a kind of graph structure stipulations method based on video features under polynary logarithm Gaussian Profile.Graph structure stipulations method based on video features under polynary logarithm Gaussian Profile, under the premise of keeping the space correlation structure of video features, construct Optical-flow Feature network graph structure, the phase information of graph structure apex Optical-flow Feature vector is contained in sequence optical flow field, and the amplitude of Optical-flow Feature vector embodies the situation of change of the target in two connected frames.It is all higher than zero in view of Optical-flow Feature vector magnitude value, the amplitude of the Optical-flow Feature vector of graph structure apex obeys polynary logarithm Gaussian Profile in video scene.The present invention proposes a kind of graph structure stipulations method of effective video features for reducing characteristic amount and capable of realizing rapid abnormal detection.
Description
Technical field
The invention belongs to the technical fields of video features optimization, and in particular to one kind is based under polynary logarithm Gaussian Profile
The graph structure stipulations method of video features.
Background technology
With the promotion of people's living standards continue to improve and population urbanization rate, public safety problem becomes increasingly conspicuous.
Under monitoring scene, anomalous event can be found and be carried out positive rescue in time, this to reduce the masses personal injury and
Property loss has great significance.Therefore, the abnormality detection based on video monitoring is then particularly important.
With the gradual extensive use increased with monitoring product of video brainpower watch and control technology, the video monitoring number of magnanimity
According to having turned into big data period more important process object.For occurring in video monitoring scene, data volume is big, data analysis
Complicated problem is handled, video monitoring system need to be improved constantly, and become more intelligent and efficient.
For the social networks of current interpersonal communication, everyone is a vertex on network, each vertex
Between have certain connection relation, a large amount of data will be generated on the vertex, constitute a complexity in this way and containing a large amount of numbers
It is believed that the network structure of breath.How by being operated to the graph structure containing important video features, to indirectly to original video
Feature is operated, and is the problem that the prior art can not capture.
Invention content
The purpose of the present invention is in view of the deficiencies of the prior art, it is proposed that one kind is based on regarding under polynary logarithm Gaussian Profile
The graph structure stipulations method of frequency feature.
The present invention is using the Optical-flow Feature in optical flow method extraction video scene, in view of the video frequency feature data of monitoring scene
It is present in the sky on the vertex of weighted network space structure and between the better expression characteristic of scramble network graph structure energy
Between relationship, the present invention describes the characteristic in video by graph structure.Before the space correlation structure for keeping video features
It puts, constructs Optical-flow Feature network graph structure, the phase information of apex Optical-flow Feature vector is contained in sequence light stream
In, the amplitude of Optical-flow Feature vector embodies the situation of change of the target in two connected frames.In view of Optical-flow Feature vector width
Angle value is all higher than zero, and the amplitude of the Optical-flow Feature vector of graph structure apex obeys polynary logarithm Gaussian Profile in video scene.
On this basis, the purpose for completing characteristic optimization is operated by the stipulations of Optical-flow Feature graph structure, the experimental results showed that this method
Data volume can effectively be reduced and can realize that rapid abnormal detects.
The concrete thought of this method：The light stream of the selection pyramid LK optical flow algorithms extraction each video frame of training set first
Feature (u, v), setting sizing grid are 24*32, choose 10*10 pixel units and extract its displacement information.Then, setting movement
Region division threshold value T takes color to encode the optical flow components higher than threshold value, the region of variation of continuous cumulative record color block
And obtain the edge contour of moving region.And using the edge contour, grid position and Optical-flow Feature of moving region come structure
Build Optical-flow Feature graph structure.Next, carrying out stipulations operation to the Optical-flow Feature graph structure.By combining Optical-flow Feature figure knot
The Downsapling method and Optical-flow Feature vector magnitude of structure Laplacian Matrix maximal eigenvector meet polynary logarithm Gaussian Profile
Characteristic amplitude mean value Downsapling method two aspect screen suitable vertex jointly, and the Gauss of thresholding is selected to weight
Kernel function connects come the inherence constructed between vertex, to carry out stipulations operation successively to feature graph structure and reach feature
The purpose of optimization.It is filtered respectively on each stipulations graph structure, and the training Optical-flow Feature after filtering is built into bag of words
Form, and be sent into the joint space-time model of LDA-iHMM, study obtains normal model parameter.The Optical-flow Feature of test set to
Amount is regarded as the input of trained normal model and is sent into abnormality detection in model, then obtains pair of each frame of test set
Number likelihood function.The log-likelihood function of test set is compared with given threshold, if the logarithm of continuous three frames test set
When likelihood function is all higher than threshold value, which is determined as exception, on the contrary then be determined as normal.
For convenience of description invention content, following term definition is done first：
Define 1：Pyramid LK optical flow algorithms
Pyramid LK optical flow algorithms are a kind of improved neighbor frame difference light splitting stream algorithm for estimating, based on following three hypothesis:
Brightness constancy, is exactly variation of the same pixel with the time, and brightness, that is, gray value will not change；Time Continuous, when
Between normally elapse, can not cause object of which movement that suddenly change occurs, the gray scale of pixel remains unchanged in region；Space one
It causes, it is also consecutive points that consecutive points, which project on image, in scene, i.e. neighbor pixel has similar variation.Define I (x) and J
(x) it is adjacent in video scene and continuous two field pictures, the gray scale of the pixel in image at any time indicates
For：
I (x)=I (x, y)
J (x)=J (x, y)
Wherein, x=[x y]TIndicate the coordinate position of a certain pixel.For pyramid LK optical flow algorithms, work as exterior light
Source stablize, it will be assumed that in the case that time interval is smaller on image the gray value of pixel be to maintain it is constant.About due to this
Beam condition, the pixel u=[u on image by finding front and back adjacent two framex uy]TWith v=[ux+dx uy+dy]TIt is right one by one
It answers, can reach the purpose of tracking.Vectorial d=[dx dy]TIndicate instantaneous velocity of the image at characteristic point u, i.e. light stream value.Gold
Word tower LK optical flow algorithm processes are to state first image pyramid, then pyramidal signature tracking, finally realize and change
For the calculating of affine light stream.
Define 2：LDA-iHMM joint space-time models
The structural mathematics expression formula of the joint space-time model of LDA-iHMM is as follows：
β | γ~GEM (γ)
In the joint space-time model of LDA-iHMM, definitionFor the Optical-flow Feature of sets of video data, wherein
The quantity that N is characterized, T represent the frame number of total video collection.The Optical-flow Feature vectorial (u, v) that we extract is to be based on bidimensional flute card
Characteristic light stream vectors (u, v), are converted under polar coordinates by X-axis and Y-axis under your coordinate system using the conversion of different coordinate between centers
Characteristic light stream vectors (ρ, θ), a length of ρ of mould and phase angle are θ.Followed by the characteristic light stream vectors (ρ, θ) under polar coordinates
Reckoning obtains word word frequency and constitutes unordered bag of words characteristic formpAnd it is z to obey themen,tMultinomial point
Cloth.In joint space-time model, theme zn,tWith state vtBetween establish undirected connection and while building sky couples.Come from some angle
It says, theme zn,tObedience state vtMultinomial distribution, utilize theme zn,tWith state vt-1Between joint transition probability matrix
Transfer between influence state.This transfer process is Di Lihe ray process, causes status number finally to tend towards stability and optimal
Number.
The present invention is based on the graph structure stipulations methods of video features under polynary logarithm Gaussian Profile, implement step
It is as follows：
Step 1：Reading resolution ratio discloses training set and test set video figure in sets of video data UMN for 240*320
Picture, setting sizing grid are 24*32, choose trained normal video frame N=180, are calculated using pyramid LK optical flow algorithms
To the movement Optical-flow Feature (u, v) of each frame.Wherein u and v is the horizontal velocity field of target movement on adjacent two video frame grid
Size and vertical velocity field size.
Step 2：One chosen in Optical-flow Feature (u, v) is compared with the threshold value T of setting, will be above threshold value T=
Optical flow components are color coded in 0.095 moving region, and the region of constantly accumulation color block finds out target moving region
Edge contour.
Step 3：Optical-flow Feature graph structure is built using the edge contour, grid position and Optical-flow Feature of moving region
G：
G={ V, ε, W }
Wherein, V is the set on vertex | V |=N, ε are undirected line sets, and W is adjacency matrix.One nonoriented edge e=(i,
J) vertex i and j are connected.
It is a diagonal matrix to spend matrix：
In formula, NiIt is the set of adjacent vertex in graph structure G.
Figure Laplacian Matrix：
L=D-W
Laplacian Matrix L real symmetric matrixes, complete orthogonal eigenvectors { ul}L=0,1 ..., N-1With non-negative feature
It is worth { λl}L=0,1 ..., N-1Between have Lul=λlul.It is after carrying out sort ascending according to characteristic value size：0=λ0<λ1≤λ2…≤
λN-1=λmax, it is λ to define the feature vector corresponding to maximum eigenvaluemax=λN-1And umax=uN-1。
Step 4：Multiple stipulations operation is carried out to the Optical-flow Feature graph structure of structure, by being pushed up on Optical-flow Feature graph structure
The reduction of point, is further reduced corresponding Optical-flow Feature quantity, to realize the optimization of characteristic information.
The screening on the vertex of feature graph structure is before this by figure Laplce's matrix L maximum feature of Optical-flow Feature graph structure
To value λmaxCorresponding maximal eigenvector umaxWith given threshold T1Vertex set V is selected compared to relatively.It is：
V1:={ i ∈ V:umax(i)≥T1}
V2:={ i ∈ V:umax(i)<T1}
Threshold value T1Before after sorting for maximal eigenvector respective valueCorresponding value, by Optical-flow Feature graph structure
Vertex set V be divided into V1And V2Two parts.And the vertex for having high weight side to be connected is retained, we retain vertex set
V1And opposite vertexes collection and V2It is rejected, the first screening on the vertex of Optical-flow Feature graph structure is completed with this.
Graph structure is screened again using the size of the amplitude equalizing value of the Optical-flow Feature vector on Optical-flow Feature graph structure vertex
Vertex.Characteristic light stream vectors (u, v) on Optical-flow Feature graph structure vertex.Convert it into Optical-flow Feature under polar coordinates to
It measures (A, θ), wherein A is the range value under polar coordinates, and θ is the phase angle under polar coordinates.
It is in a dimension for having the feature graph structure on n vertex, the feature vector amplitude A on each vertex
Stochastic variable.N is tieed up graph structure signal x=(x by us1,...,xN)TStipulations are that n-k ties up graph structure signal x1=(x1,...,
xN-K)T, because not knowing that it is best that select which k vertex, goes to screening vertex from the angle of probability, makes information
Loss difference is reduced to minimum.Because the distribution of the Optical-flow Feature vector magnitude on feature graph structure vertex is polynary logarithm Gauss
Distribution, so the difference entropy after stipulations is allowed to reach maximum value, is shown below：
In above formula object function, e is the nature truth of a matter, D x1Covariance matrix, μiFor each vertex Optical-flow Feature to
The amplitude equalizing value of amount.In order to make it is sized after information entropy maximization, select best x1The maximum mutual information between x is found,
Allow the amplitude equalizing value μ for leaving light stream feature vector on vertexiIt is big as far as possible.In view of the number of vertex mistake of graph structure in video
It is more, and the characteristic signal on vertex is very small, so the value of its determinant would tend to infinitesimal, therefore seeks difference
The maximum value of entropy then ignores this.Therefore, the screening process again on graph structure vertex is in screening vertex set V for the first time1Basis
On, utilize amplitude equalizing value and the given threshold T of light stream vectors on vertex2After screened again, obtain V11。
V11:={ i ∈ V1:μi≥T2}
V12:={ i ∈ V1:μi<T2}
Threshold size T in upper two formula2For optical flow field amplitude equalizing value μ on each vertex of graph structureiBy what is arranged from big to small
BeforeCorresponding mean value size, by graph structure vertex set V1It is divided into V11And V12Two parts, and retain vertex set V11And
Reject vertex set V12.After vertex twice is screened, graph structure number of vertex reduces half, and overall pending data volume also subtracts
Few half.
The vertex set V that will be filtered out11The inherence between vertex is constructed using the gaussian kernel function of a thresholding even
It connects, defines the side right weight W of nonoriented edge e=(i, j) connection vertex an i and ji,j, constitute sized subgraph structure G={ V11,ε,
W }, wherein V11It is the set on vertex after screening, ε is undirected line set, and W is the adjacency matrix containing weight.
Wherein κ is threshold value, parameterN is graph structure vertex sum, and dist (i, j) is
Euclidean distance between vertex i and j, threshold value κ=0.6.
Step 5：On the graph structure of stipulations successively, the Optical-flow Feature of training set and test set to feeding carries out respectively
It filters and builds corresponding Feature Words bag form.
Step 6：By the parameter of bag of words feeding LDA-iHMM joint space-time models middle school's acquistion of training set feature to model
And log-likelihood function of the training set per frame, finally again using the Optical-flow Feature vector in the bag of words of test set as having trained " just
Often " input of model carries out the abnormality detection of video, obtains the log-likelihood function of each frame of test set.
Step 7：The log-likelihood function of test set compares with given threshold, if pair of continuous three frames test set
On the contrary number likelihood functions are when being all higher than threshold value Th, which is determined as exception, then be determined as normally；It is regarded until entirely testing
Frequency, which collects all detections, to be completed.
The advantage of the invention is that：
The method that the present invention mainly proposes the graph structure stipulations of video features under polynary logarithm Gaussian Profile is applied to regard
The optimization processing of frequency feature.On this basis, the purpose for completing characteristic optimization is operated by the stipulations of Optical-flow Feature graph structure.
By going to verify the applicability of the algorithm with the simulation result of UMN data set abnormality detections, the experimental results showed that this method can be with
It effectively reduces data volume and the calculating speed of video Outlier Detection Algorithm can be improved.
Description of the drawings
Fig. 1 is the flow chart of the present invention.
Fig. 2 is the joint space-time model of LDA-iHMM.
Specific implementation mode
The implementing procedure figure of the present invention is as shown in Figure 1, specific implementation step is as follows：
Step 1：Reading resolution ratio discloses training set and test set video figure in sets of video data UMN for 240*320
Picture, setting sizing grid are 24*32, choose trained normal video frame N=180, are calculated using pyramid LK optical flow algorithms
To the movement Optical-flow Feature (u, v) of each frame.Wherein u and v is the horizontal velocity field of target movement on adjacent two video frame grid
Size and vertical velocity field size.
Step 2：One chosen in Optical-flow Feature (u, v) is compared with the threshold value T of setting, will be above threshold value T=
Optical flow components are color coded in 0.095 moving region, and constantly accumulation color block region finds out target moving region
Edge contour.
Step 3：Optical-flow Feature graph structure is built using the edge contour, grid position and Optical-flow Feature of moving region
G：
G={ V, ε, W }
Wherein, V is the set on vertex | V |=N, ε are undirected line sets, and W is adjacency matrix.One nonoriented edge e=(i,
J) vertex i and j are connected.
It is a diagonal matrix to spend matrix：
In formula, NiIt is the set of adjacent vertex in graph structure G.
Figure Laplacian Matrix：
L=D-W
Laplacian Matrix L real symmetric matrixes, complete orthogonal eigenvectors { ul}L=0,1 ..., N-1With non-negative feature
It is worth { λl}L=0,1 ..., N-1Between have Lul=λlul.It is after carrying out sort ascending according to characteristic value size：0=λ0<λ1≤λ2…≤
λN-1=λmax, it is λ to define the feature vector corresponding to maximum eigenvaluemax=λN-1And umax=uN-1。
Step 4：Multiple stipulations operation is carried out to the Optical-flow Feature graph structure of structure, by being pushed up on Optical-flow Feature graph structure
The reduction of point, is further reduced corresponding Optical-flow Feature quantity, to realize the optimization of characteristic information.
The screening on the vertex of feature graph structure is before this by figure Laplce's matrix L maximum feature of Optical-flow Feature graph structure
To value λmaxCorresponding maximal eigenvector umaxWith given threshold T1Vertex set V is selected compared to relatively.It is：
V1:={ i ∈ V:umax(i)≥T1}
V2:={ i ∈ V:umax(i)<T1}
Threshold value T1Before after sorting for maximal eigenvector respective valueCorresponding value, by Optical-flow Feature graph structure
Vertex set V be divided into V1And V2Two parts.And the vertex for having high weight side to be connected is retained, we retain vertex set
V1And opposite vertexes collection and V2It is rejected, the first screening on the vertex of Optical-flow Feature graph structure is completed with this.
Graph structure is screened again using the size of the amplitude equalizing value of the Optical-flow Feature vector on Optical-flow Feature graph structure vertex
Vertex.Characteristic light stream vectors (u, v) on Optical-flow Feature graph structure vertex.Convert it into Optical-flow Feature under polar coordinates to
It measures (A, θ), wherein A is the range value under polar coordinates, and θ is the phase angle under polar coordinates.
It is in a dimension for having the feature graph structure on n vertex, the feature vector amplitude A on each vertex
Stochastic variable.N is tieed up graph structure signal x=(x by us1,...,xN)TStipulations are that n-k ties up graph structure signal x1=(x1,...,
xN-K)T, because not knowing that it is best that select which k vertex, goes to screening vertex from the angle of probability, makes information
Loss difference is reduced to minimum.Because the distribution of the Optical-flow Feature vector magnitude on feature graph structure vertex is polynary logarithm Gauss
Distribution, so we allow the difference entropy after stipulations to reach maximum value, is shown below：
In above formula object function, e is the nature truth of a matter, D x1Covariance matrix, μiFor each vertex Optical-flow Feature to
The amplitude equalizing value of amount.In order to make it is sized after information entropy maximization, select best x1The maximum mutual information between x is found,
Allow the amplitude equalizing value μ for leaving light stream feature vector on vertexiIt is big as far as possible.In view of the number of vertex mistake of graph structure in video
It is more, and the characteristic signal on vertex is very small, so the value of its determinant would tend to infinitesimal, therefore seeks difference
The maximum value of entropy then ignores this.Therefore, the screening process again on graph structure vertex is in screening vertex set V for the first time1Basis
On, utilize amplitude equalizing value and the given threshold T of light stream vectors on vertex2After screened again, obtain V11。
V11:={ i ∈ V1:μi≥T2}
V12:={ i ∈ V1:μi<T2}
Threshold size T in upper two formula2For optical flow field amplitude equalizing value μ on each vertex of graph structureiBy what is arranged from big to small
BeforeCorresponding mean value size, by graph structure vertex set V1It is divided into V11And V12Two parts, and retain vertex set V11And
Reject vertex set V12.After vertex twice is screened, graph structure number of vertex reduces half, and overall pending data volume also subtracts
Few half.
The vertex set V that will be filtered out11The inherence between vertex is constructed using the gaussian kernel function of a thresholding even
It connects, defines the side right weight W of nonoriented edge e=(i, j) connection vertex an i and ji,j, constitute sized subgraph structure G={ V11,ε,
W }, wherein V11It is the set on vertex after screening, ε is undirected line set, and W is the adjacency matrix containing weight.
Wherein κ is threshold value, parameterN is graph structure vertex sum, and dist (i, j) is
Euclidean distance between vertex i and j, threshold value κ=0.6.
Step 5：On the graph structure of stipulations successively, the Optical-flow Feature of training set and test set to feeding carries out respectively
It filters and builds corresponding Feature Words bag form.
Step 6：By the parameter of bag of words feeding LDA-iHMM joint space-time models middle school's acquistion of training set feature to model
And log-likelihood function of the training set per frame, the joint space-time model of LDA-iHMM is as shown in Fig. 2, last again by test set
Optical-flow Feature vector in bag of words obtains test set as the abnormality detection for having trained the input of " normal " model to carry out video
The log-likelihood function of each frame.
Step 7：The log-likelihood function of test set compares with given threshold, if pair of continuous three frames test set
On the contrary number likelihood functions are when being all higher than threshold value Th, which is determined as exception, then be determined as normally；It is regarded until entirely testing
Frequency, which collects all detections, to be completed.
Claims (5)
1. the graph structure stipulations method based on video features under polynary logarithm Gaussian Profile, it is characterised in that include the following steps：
Step 1：The video for reading in the training set and test set in open sets of video data UMN, selects trained normal video frame, if
Sizing grid is set, video consolidated movement Optical-flow Feature (u, v) is found out using pyramid LK optical flow algorithms；Wherein, u regards for adjacent two
The horizontal velocity field size that target moves on frequency frame grid, v are the vertical velocity field of target movement on adjacent two video frame grid
Size；
Step 2：It chooses Optical-flow Feature (u, v) to be color coded higher than the optical flow components for dividing movement threshold T, and constantly accumulates
It records the region of color block and obtains the edge contour of moving region；
Step 3：Optical-flow Feature graph structure G is built using the edge contour, grid position and Optical-flow Feature of moving region；
Step 4：Multiple stipulations operation is carried out to the Optical-flow Feature graph structure of structure, is subtracted by vertex on Optical-flow Feature graph structure
It is few, it is further reduced corresponding Optical-flow Feature quantity, to realize the optimization of characteristic information；
Step 5：On the graph structure of stipulations successively, the Optical-flow Feature of training set and test set to feeding is filtered simultaneously respectively
Build corresponding Feature Words bag form；
Step 6：By the parameter and instruction of bag of words feeding LDA-iHMM joint space-time models middle school's acquistion of training set feature to model
Practice log-likelihood function of the collection per frame, finally again using the Optical-flow Feature vector in the bag of words of test set as having trained " normal " mould
The input of type carries out the abnormality detection of video, obtains the log-likelihood function of each frame of test set；
Step 7：The log-likelihood function of test set compares with given threshold, if the log-likelihood of continuous three frames test set
When function is all higher than threshold value Th, which is determined as exception, on the contrary then be determined as normal；Until entire test video collection whole
Detection is completed.
2. the graph structure stipulations method according to claim 1 based on video features under polynary logarithm Gaussian Profile, special
Sign is that the resolution ratio of training set and test set in the open video set data UNM that step 1 is read in is 240*320, chooses instruction
Experienced preceding N normal video frames, N=180；Setting sizing grid is 24*32.
3. the graph structure stipulations method according to claim 1 based on video features under polynary logarithm Gaussian Profile, special
Sign, which is to set in step 2, divides movement threshold T as 0.095.
4. the graph structure stipulations method according to claim 1 based on video features under polynary logarithm Gaussian Profile, special
Sign is the Optical-flow Feature graph structure G described in step 3：
G={ V, ε, W }
Wherein, V is the set on vertex | V |=N, ε are undirected line sets, and W is adjacency matrix, nonoriented edge e=(i, a j) connection
Vertex i and j；
It is a diagonal matrix to spend matrix：
In formula, NiIt is the set of adjacent vertex in graph structure G；
The Laplacian Matrix of figure：
L=D-W
Laplacian Matrix L is symmetrical matrix, complete orthogonal eigenvectors { ul}L=0,1 ..., N-1With non-negative characteristic value
{λl}L=0,1 ..., N-1Between have Lul=λlul, it is after carrying out sort ascending according to characteristic value size：0=λ0<λ1≤λ2…≤λN-1=
λmax, it is λ to define the feature vector corresponding to maximum eigenvaluemax=λN-1And umax=uN-1。
5. the graph structure stipulations method according to claim 1 based on video features under polynary logarithm Gaussian Profile, special
Sign be step 4 the specific implementation process is as follows：
The screening on the vertex of feature graph structure is before this from the figure Laplce matrix L maximum feature of Optical-flow Feature graph structure to value
λmaxCorresponding maximal eigenvector umaxWith given threshold T1Vertex set V is selected compared to relatively, is：
V1:={ i ∈ V:umax(i)≥T1}
V2:={ i ∈ V:umax(i)<T1}
Threshold value T1Before after sorting for maximal eigenvector respective valueCorresponding value, by the vertex of Optical-flow Feature graph structure
Collection V is divided into V1And V2Two parts, and the vertex for having high weight side to be connected is retained, retain vertex set V1And to top
Point set and V2It is rejected, the first screening on the vertex of Optical-flow Feature graph structure is completed with this；
Graph structure vertex is screened again using the size of the amplitude equalizing value of the Optical-flow Feature vector on Optical-flow Feature graph structure vertex,
Characteristic light stream vectors (u, v) on Optical-flow Feature graph structure vertex, convert it under polar coordinates Optical-flow Feature vector (A,
θ), wherein A is the range value under polar coordinates, and θ is the phase angle under polar coordinates；
It is the random change in a dimension for having the feature graph structure on n vertex, the feature vector amplitude A on each vertex
Amount, by n dimension graph structure signal x=(x1,...,xN)TStipulations are that n-k ties up graph structure signal x1=(x1,...,xN-K)T；After stipulations
Difference entropy reach maximum value, be shown below：
In above formula object function, e is the nature truth of a matter, D x1Covariance matrix, μiFor each vertex Optical-flow Feature vector
Amplitude equalizing value；The screening process again on graph structure vertex is in screening vertex set V for the first time1On the basis of, utilize light stream on vertex
The amplitude equalizing value of vector and given threshold T2The vertex for being screened Optical-flow Feature graph structure more again, obtains V11：
V11:={ i ∈ V1:μi≥T2}
V12:={ i ∈ V1:μi<T2}
Threshold size T in upper two formula2For optical flow field amplitude equalizing value μ on each vertex of graph structureiBefore arranging from big to smallCorresponding mean value size, by graph structure vertex set V1It is divided into V11And V12Two parts, and retain vertex set V11And it picks
Except vertex set V12；
The vertex set V that will be filtered out11The inherent connection between vertex, definition are constructed using the gaussian kernel function of a thresholding
The side right weight W of one nonoriented edge e=(i, j) connection vertex i and ji,j, constitute sized subgraph structure G={ V11, ε, W }, wherein
V11It is the set on vertex after screening, ε is undirected line set, and W is the adjacency matrix containing weight.
Wherein κ is threshold value, parameterN is graph structure vertex sum, and dist (i, j) is vertex i
Euclidean distance between j, threshold value κ=0.6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810048588.6A CN108509834B (en) | 2018-01-18 | 2018-01-18 | Graph structure reduction method based on video features under multivariate logarithmic Gaussian distribution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810048588.6A CN108509834B (en) | 2018-01-18 | 2018-01-18 | Graph structure reduction method based on video features under multivariate logarithmic Gaussian distribution |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108509834A true CN108509834A (en) | 2018-09-07 |
| CN108509834B CN108509834B (en) | 2022-05-31 |
Family
ID=63374896
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810048588.6A Active CN108509834B (en) | 2018-01-18 | 2018-01-18 | Graph structure reduction method based on video features under multivariate logarithmic Gaussian distribution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108509834B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507277A (en) * | 2020-04-20 | 2020-08-07 | 国网安徽省电力有限公司 | Construction method and application of operation flow normative recognition model |
| CN111614965A (en) * | 2020-05-07 | 2020-09-01 | 武汉大学 | Unmanned aerial vehicle video image stabilization method and system based on image grid optical flow filtering |
| CN113239039A (en) * | 2021-05-11 | 2021-08-10 | 北京理工新源信息科技有限公司 | Dynamic data storage method, query method, management method and management system |
| WO2022193507A1 (en) * | 2021-03-15 | 2022-09-22 | 深圳市慧鲤科技有限公司 | Image processing method and apparatus, device, storage medium, program, and program product |
| CN116433887A (en) * | 2023-06-12 | 2023-07-14 | 山东鼎一建设有限公司 | Building rapid positioning method based on artificial intelligence |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103530603A (en) * | 2013-09-24 | 2014-01-22 | 杭州电子科技大学 | Video abnormality detection method based on causal loop diagram model |
| CN104820824A (en) * | 2015-04-23 | 2015-08-05 | 南京邮电大学 | Local abnormal behavior detection method based on optical flow and space-time gradient |
| CN105469054A (en) * | 2015-11-25 | 2016-04-06 | 天津光电高斯通信工程技术股份有限公司 | Model construction method of normal behaviors and detection method of abnormal behaviors |
| CN105787472A (en) * | 2016-03-28 | 2016-07-20 | 电子科技大学 | Abnormal behavior detection method based on time-space Laplacian Eigenmaps learning |
| CN106407327A (en) * | 2016-08-31 | 2017-02-15 | 广州精点计算机科技有限公司 | Similar image searching method and device based on HOG (Histogram of Oriented Gradient) and visual word bag |
| CN106548153A (en) * | 2016-10-27 | 2017-03-29 | 杭州电子科技大学 | Video abnormality detection method based on graph structure under multi-scale transform |
-
2018
- 2018-01-18 CN CN201810048588.6A patent/CN108509834B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103530603A (en) * | 2013-09-24 | 2014-01-22 | 杭州电子科技大学 | Video abnormality detection method based on causal loop diagram model |
| CN104820824A (en) * | 2015-04-23 | 2015-08-05 | 南京邮电大学 | Local abnormal behavior detection method based on optical flow and space-time gradient |
| CN105469054A (en) * | 2015-11-25 | 2016-04-06 | 天津光电高斯通信工程技术股份有限公司 | Model construction method of normal behaviors and detection method of abnormal behaviors |
| CN105787472A (en) * | 2016-03-28 | 2016-07-20 | 电子科技大学 | Abnormal behavior detection method based on time-space Laplacian Eigenmaps learning |
| CN106407327A (en) * | 2016-08-31 | 2017-02-15 | 广州精点计算机科技有限公司 | Similar image searching method and device based on HOG (Histogram of Oriented Gradient) and visual word bag |
| CN106548153A (en) * | 2016-10-27 | 2017-03-29 | 杭州电子科技大学 | Video abnormality detection method based on graph structure under multi-scale transform |
Non-Patent Citations (3)
| Title |
|---|
| CHA ZHANG 等: "Graph Signal Processing – A Probabilistic Framework", 《COMPUTER SCIENCE》 * |
| DAVID I SHUMAN 等: "A Multiscale Pyramid Transform for Graph Signals", 《IEEE TRANSACTIONS ON SIGNAL PROCESSING》 * |
| 汪洪流 等: "图结构多尺度变换的视频异常检测", 《中国图象图形学报》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507277A (en) * | 2020-04-20 | 2020-08-07 | 国网安徽省电力有限公司 | Construction method and application of operation flow normative recognition model |
| CN111507277B (en) * | 2020-04-20 | 2023-05-09 | 国网安徽省电力有限公司 | Construction method and application of workflow normalization recognition model |
| CN111614965A (en) * | 2020-05-07 | 2020-09-01 | 武汉大学 | Unmanned aerial vehicle video image stabilization method and system based on image grid optical flow filtering |
| CN111614965B (en) * | 2020-05-07 | 2022-02-01 | 武汉大学 | Unmanned aerial vehicle video image stabilization method and system based on image grid optical flow filtering |
| WO2022193507A1 (en) * | 2021-03-15 | 2022-09-22 | 深圳市慧鲤科技有限公司 | Image processing method and apparatus, device, storage medium, program, and program product |
| CN113239039A (en) * | 2021-05-11 | 2021-08-10 | 北京理工新源信息科技有限公司 | Dynamic data storage method, query method, management method and management system |
| CN113239039B (en) * | 2021-05-11 | 2023-11-14 | 北京理工新源信息科技有限公司 | Dynamic data storage method, query method, management method and management system |
| CN116433887A (en) * | 2023-06-12 | 2023-07-14 | 山东鼎一建设有限公司 | Building rapid positioning method based on artificial intelligence |
| CN116433887B (en) * | 2023-06-12 | 2023-08-15 | 山东鼎一建设有限公司 | Building rapid positioning method based on artificial intelligence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108509834B (en) | 2022-05-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108509834A (en) | Graph structure stipulations method based on video features under polynary logarithm Gaussian Profile | |
| Liu et al. | Crowd counting with deep structured scale integration network | |
| Tao et al. | Smoke detection based on deep convolutional neural networks | |
| CN108830252B (en) | Convolutional neural network human body action recognition method fusing global space-time characteristics | |
| Cheng et al. | Improving the learning of multi-column convolutional neural network for crowd counting | |
| CN106548153B (en) | Video abnormality detection method based on graph structure under multi-scale transform | |
| CN106997597B (en) | It is a kind of based on have supervision conspicuousness detection method for tracking target | |
| CN104504394B (en) | A kind of intensive Population size estimation method and system based on multi-feature fusion | |
| CN110119703A (en) | The human motion recognition method of attention mechanism and space-time diagram convolutional neural networks is merged under a kind of security protection scene | |
| CN103839065B (en) | Extraction method for dynamic crowd gathering characteristics | |
| CN107169415A (en) | Human motion recognition method based on convolutional neural networks feature coding | |
| CN101470809B (en) | Moving object detection method based on expansion mixed gauss model | |
| CN111582068A (en) | Method for detecting wearing state of personal mask | |
| CN104182772A (en) | Gesture recognition method based on deep learning | |
| CN107633226A (en) | A kind of human action Tracking Recognition method and system | |
| CN108960047B (en) | Face duplication removing method in video monitoring based on depth secondary tree | |
| CN108470077A (en) | A kind of video key frame extracting method, system and equipment and storage medium | |
| CN107066963B (en) | A kind of adaptive people counting method | |
| CN111639577A (en) | Method for detecting human faces of multiple persons and recognizing expressions of multiple persons through monitoring video | |
| CN104376334A (en) | Pedestrian comparison method based on multi-scale feature fusion | |
| CN107665361A (en) | A kind of passenger flow counting method based on recognition of face | |
| CN114758288A (en) | Power distribution network engineering safety control detection method and device | |
| WO2021139171A1 (en) | Facial enhancement based recognition method, apparatus and device, and storage medium | |
| CN109902613A (en) | A kind of human body feature extraction method based on transfer learning and image enhancement | |
| CN104537381B (en) | A kind of fuzzy image recognition method based on fuzzy invariant features |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |